Relative clause attachment in Mandarin Chinese: insights from classifier-noun agreement

Introduction: Previous studies have shown that relative clause (RC) attachment preferences vary across languages, often influenced by factors like morphosyntactic agreement (e.g., number and gender). Mandarin Chinese, with its limited inflectional morphemes compared to Indo-European languages, provides a distinct context for examining this. This study explores relative clause attachment ambiguity in Mandarin by manipulating classifier-noun agreement.

Method: This study conducted two self-paced reading experiments to investigate the influence of an initial classifier on comprehenders' anticipation of its associated noun and the impact of this prediction on RC attachment preferences in Mandarin Chinese.

Results: Experiment 1 revealed a significant effect of classifier-noun agreement in offline comprehension: there was an increase in selecting the high-attachment noun (NP_high) as the RC attachment site when the classifier agreed with NP_high, whereas there was a decrease in selecting NP_high when the classifier agreed with the low-attachment noun (NP_low). Online processing results supported this effect, showing that classifiers guide comprehenders' expectations by pre-activating semantic features of the upcoming noun, thus modulating RC attachment preferences. Experiment 2 introduced semantic compatibility between the RC and potential attachment nouns as an additional disambiguating cue, revealing a reliable prediction effect for the upcoming noun. Although the classifier's prediction effect was diminished, it remained influential in this condition.

Discussion: This study highlights the complexity of relative clause attachment in Mandarin, demonstrating the significant predictive roles of classifier-noun agreement and semantic compatibility.

1 Introduction

The phenomenon of relative clause attachment ambiguity has attracted considerable attention among researchers. In English, resolving the attachment of relative clauses (RCs) is challenging when the head noun is inside a complex nominal phrase consisting of two noun phrases (NP_high of NP_low). This complexity introduces ambiguity, as the relative clause can modify either NP_high or NP_low (Cuetos and Mitchell, 1988).

Example (1) illustrates the attachment ambiguity with this type of sentence. When a relative clause who was on the balcony follows a complex noun phrase the servant of the actress, it can typically be attached to two potential noun sites. Thus, example (1) could imply that either the servant or the actress was the person on the balcony. To disambiguate the sentence, it becomes necessary to determine whether the relative clause attaches to the first potential host site (NP_high, the servant) or the second (NP_low, the actress). Resolutions of the first kind will be referred to as high attachment (HA) as NP_high occupies a higher position in the syntactic structure. Correspondingly, instances of the second kind which favor NP_low will be labeled low attachment (LA) (see Figure 1 for illustrations).

(1) Someone shot the servant of the actress who was on the balcony.

Figure 1

Figure 1. Relative clause attachment preference structures in English [high-attachment: the upper (A); low-attachment: the lower (B)].

1.1 Cross-linguistic differences in relative clause attachments

Cross-linguistic studies have revealed variation in attachment preferences across languages. For instance, languages such as Spanish, Dutch, Italian, German, French, Japanese, and Korean tend to favor high attachment resolutions (Cuetos and Mitchell, 1988; de Vincenzi and Job, 1993; Hemforth et al., 1994; Brysbaert and Mitchell, 1996; Kamide and Mitchell, 1997; Zagar et al., 1997; Fernández, 2002; Lee, 2021). On the other hand, languages such as English and Mandarin tend to exhibit a preference for low attachment (Cuetos and Mitchell, 1988; Fernández, 2002; Kwon et al., 2019).

However, the distinctions between the two language groups are not as clear-cut as they seemed. Research findings often yield contradictory results due to methodological differences. For example, paper-based questionnaires typically tap into offline processing (later stages) while self-paced reading or eye-tracking techniques tap into online processing (earlier stages). Consequently, studies utilizing offline questionnaire tasks consistently report a high-attachment preference in Spanish (Cuetos and Mitchell, 1988; Fernández, 2002). In contrast, studies employing self-paced reading tasks have yielded inconsistent results, finding either a low-attachment preference (Fernández, 2002) or a high-attachment preference (Mitchell and Cuetos, 1991). Moreover, the semantic and length properties of the test materials, such as using human versus non-human nouns or the length and complexity of the relative clauses, also contribute to these differences (Fernández, 2002; Hemforth et al., 2015; Kwon et al., 2019).

Mitchell and colleagues conducted a series of studies investigating relative clause attachment in several languages. One of the pioneering studies by Cuetos and Mitchell (1988) compared RC attachment preferences in Spanish and English. Despite sharing the SVO word order and head-initial RC structure, Spanish and English differ in the word order of adjectives and nouns. Unlike the prenominal adjective order in English, Spanish commonly follows a postnominal adjective order, where the adjective appears after the noun it modifies (Cuetos and Mitchell, 1988). Cuetos and Mitchell (1988) used equivalent materials, providing literal translations of sentences in both English and Spanish, in an offline questionnaire task (as shown in example (1) above). In this task, each sentence was followed by a question, such as “Who was on the balcony?” to determine the attachment preference. The results revealed a significant tendency among Spanish participants to attach the RC to the first noun (NP_high, i.e., servant), showing a strong high-attachment preference, whereas English participants preferred attachment to the second noun (NP_low, i.e., actress), indicating a low-attachment preference. This finding demonstrated a cross-linguistic difference in attachment preferences between Spanish and English. Subsequent studies, such as Fernández (2002), confirmed this cross-linguistic distinction, with questionnaire results replicating the findings in Cuetos and Mitchell (1988). In Fernández's (2002) study, Spanish readers exhibited a higher likelihood (57%) of choosing the high-attachment site (NP_high) compared to English readers (43%).

Mandarin, on the other hand, follows an SVO pattern and a head-final RC structure. In addition, Mandarin employs the possessive particle de to express possession or genitive relationships instead of using a Norman genitive construction. This particle, de, is positioned after the possessor noun to indicate ownership. For instance, corresponding to the servant of the actress, nvyanyuan “actress” de “DE” puren “servant” in Mandarin would mean the actress' servant, as shown in example (2) below. Therefore, in Mandarin, the first NP (actress) serves as the low attachment site, while the second NP (servant) serves as the high attachment site. Kwon et al. (2019) found a prevailing preference toward low attachment in Mandarin in an offline sentence completion task.

(2) Mandarin (adapted from Shen, 2006)

Mouren    dasile         zhanzai       nvyanyuan   puren.

                       yangtai-      de

                        shang de

Someone-    shot           on the          actress-      servant.

Nom                  balcony-      particle

                    Relativizer

Subject_      Predicate_ RC            NP1-Low     NP2-

matrix       matrix                  High

“Someone shot the servant of the actress who was on the balcony.”

As we can see, relative clauses involving global ambiguity have been widely used in offline tasks to explore attachment preferences in a variety of languages. Cross-linguistic differences have been observed in languages that differ with respect to typological properties such as word order variability, alternative forms for expressing genitive meanings, the overt/covert relative pronouns, as well as left-branching/right branching RC structures.

1.2 Factors that modulate relative clause attachment preference

Most studies (e.g., Lee and Kweon, 2004; Kwon et al., 2019) discussed readers' attachment preferences within the context of offline tasks, tapping into the later phases of processing. To gain insights into readers' preferences during the earliest phases of processing, a number of studies have employed online tasks, such as self-paced reading and eye-tracking techniques, measuring differences in terms of processing difficulty associated with forcing attachment in either direction. For instance, an online task may compare reading times across three conditions: ambiguous, forced high-attachment, and forced low-attachment (Fernández, 2002). Readers' preferences in the early processing phases would manifest as increased reading times at critical regions when encountering materials that contain the dis-preferred forced attachment, as compared to reading times associated with either ambiguous materials or those containing the preferred forced attachment (Fernández, 2002). Studies on Indo-European languages often employ morphosyntactic agreement as a means to disambiguate relative clauses with complex head nouns. English, for instance, always utilizes number agreement to force high or low attachments, as shown in example (3) (Fernández, 2002, p. 195).

(3) Number agreement in English

 a.... the nephew of the teacher that was...  (Ambiguous)

 b.... the nephew of the teachers that was...  (Forced HA)

 c.... the nephews of the teacher that was...  (Forced LA)

In terms of early attachment preferences, Fernández (2002) conducted a self-paced reading task comparing reading times in forced low and forced high conditions in both English and Spanish. Surprisingly, no differences were found between the two languages, as both exhibited faster reading times in the forced low attachment condition (3c) compared to the forced high attachment condition (3b). These results suggest that the initial attachment choice in relative clauses tends to be the low site. Fernández attributed the observed cross-language differences between Spanish and English in offline tasks to the influence of post-syntactic information. Similarly, a study by de Vincenzi and Job (1993) employed gender agreement in Italian to disambiguate relative clause attachment. Contrary to the high-attachment preference found in offline questionnaires, a self-paced reading task revealed a low attachment preference in the early phase, suggesting that the initial low-attachment choice is subsequently adjusted to high-attachment considering post-syntactic information. Kamide and Mitchell (1997) conducted a study in Japanese where they manipulated the pragmatic compatibility between the RC and the two potential attachment sites, aiming to force the RC to attach to either NP. Using a self-paced reading task with seven segmented regions (RC/NP1-Low/Gen/NP2-High/Acc/Matrix subject/Final), the study revealed significantly longer reading times in the regions associated with NP1-Low and Gen in the forced high-attachment condition, compared to the forced low-attachment condition. Interestingly, the Final region exhibited significantly longer reading times in the forced low-attachment condition than in the forced high-attachment condition. These findings suggest that parsers initially exhibit a low-attachment when encountering the first potential attachment site (NP1-Low), but their preference switches to high attachment by the end of the sentence.

In contrast to the findings observed in Japanese, the languages Korean and Mandarin exhibit consistent attachment preferences in both offline and online tasks. Specifically, Korean consistently shows high attachment as the initial and eventual choice, while Mandarin consistently favors low attachment. Lee and Kweon (2004) employed a disambiguating device in the form of subject-verb agreement with an honorific marker -si. By manipulating the agreement between the verb and the potential agents using the honorific marker, the relative clauses are forced to attach to either site. Similarly, Shen (2006) manipulated the semantic compatibility of the RC with alternative host sites and conducted both a self-paced reading task and an offline questionnaire. The results from Shen (2006) demonstrated a low-attachment preference that persisted from the late online processing phase into the offline phase.

While online tasks have commonly utilized morphosyntactic agreement (e.g., number and gender) in relative clauses to disambiguate attachment preferences in languages like English, Spanish, Italian, and French, Mandarin Chinese differs from most Indo-European languages in having a relatively limited set of inflectional morphemes. However, Mandarin features classifier-noun agreement, which remains largely unexplored in relation to the RC attachment preferences. The present study addressed this issue by conducting two self-paced reading experiments using classifier-noun agreement as a disambiguating cue to resolve relative clause attachment preferences.

1.3 Predictions in sentences in relation to classifier-noun agreement

1.3.1 Prediction in language processing

Language comprehenders actively make predictions about upcoming linguistic units, whether they are syllables, words, or grammatical structures. The ability to incrementally form hypotheses about incoming linguistic material, known as predictive processing, is a fundamental property of sentence processing (Kuperberg and Jaeger, 2016; Pickering and Gambi, 2018). Comprehenders use context to make predictions, pre-activating linguistic representations before actually encountering them (Pickering and Garrod, 2013; Pickering and Gambi, 2018). This pre-activation allows the brain to prepare for expected linguistic input in advance, facilitating quicker and more efficient processing.

A number of studies have demonstrated that people can predict semantic information of upcoming words, particularly in highly constraining contexts (Federmeier and Kutas, 1999; Kamide et al., 2003; Wang et al., 2018). However, less is known about predictions in less-constraining contexts, which may allow for the prediction of general information characterizing a group of words rather than specific words (Huang et al., 2023). Some studies have explored this issue by using morphosyntactic dependencies triggered by cues such as gender-marking (German: Hopp and Lemmerth, 2018), case marking (Turkish: Özge et al., 2019; German: Özge et al., 2022), and honorific marking (Lee and Yoo, 2023). These studies have found that people can use morphosyntax to generate predictions during the course of processing.

Hopp and Lemmerth (2018), for example, examined whether gender agreement between pre-nominal determiners/adjectives and nouns in German can aid comprehenders in predicting the upcoming nouns. Using a visual-world task, they found that gender marking on the determiner or adjective helped participants identify the object among competitors. Participants spent less time identifying the object when the gender marking (masculine or feminine) on the determiner or adjective provided disambiguating cues compared to when the gender marking (neuter) provided no disambiguating clues. These findings suggest that gender markings on determiners or adjectives pre-activate the gender information of the upcoming nouns, facilitating faster and more accurate identification of objects that share the gender features with the predicted nouns.

Lee and Yoo (2023) also demonstrated the pre-activation effect of honorific agreement between the subject and verb in Korean. In a self-paced reading task, they manipulated the honorific agreement by varying whether the subjects had honorific features. Results showed that subjects with honorific features elicited longer reading times when the verbs do not carry an honorific marker, compared to subjects without honorific features, supporting predictive processing, as participants pre-activate the honorific information when they encounter a subject with honorific features and expect an honorific marker -si on the following verb before encountering it. In addition, Özge et al. (2019, 2022) found that both Turkish and German children as young as four years old were able to use sentence-initial case marking cues to predict the thematic role of the upcoming argument.

These studies have demonstrated the role of morphosyntactic agreement and dependencies in facilitating predictive processing in less-constraining contexts, suggesting that comprehenders use morphosyntactic cues to anticipate and prepare for upcoming linguistic information. However, gender, case, and honorific agreements are highly grammaticalized and correlate with overt morpho-syntactic markers. This raises the question of whether comprehenders can also rely on agreement or dependency relationships that are not explicitly marked to predict upcoming linguistic information during sentence comprehension. To explore this, the present study adopted the classifier-noun agreement in Mandarin Chinese to test whether people can utilize the semantic constraints of classifiers to make predictions for the upcoming nouns.

1.3.2 Classifier-noun agreement and its prediction effect

Classifiers are systems used to mark and categorize noun classes (Erbaugh, 2006; Ahrens and Huang, 2016). Mandarin is one of the languages that employs numeral classifiers, which appear after a number or determiner and before a noun (see example (4) and Figure 2 for illustrations). In Mandarin, classifiers exhibit selectivity for specific noun classes, establishing a classifier-noun agreement through selectional restrictions. For instance, the classifier tiao “CL(objects/animals)” typically pairs with objects or animals that are long, thin, cylindrical, and flexible, such as yu “fish” and maojin “towel”, while it cannot be used with other nouns like flowers or books (Ahrens and Huang, 2016, p. 177).

(4) Zhe    san     tiao    yu.

 DET  NUM  CL    N

 This      three   CL    fish

 ‘these    three      fishes'

Figure 2

Figure 2. A structural representation of a numeral classifier phrase in Mandarin.

Several studies (Zhou et al., 2010; Chou et al., 2015; Chan, 2019) have investigated the dependencies between classifiers and nouns. Zhou et al. (2010) conducted an event-related potential (ERP) study to explore the temporal neural dynamics of semantic integration processes in a hierarchical structure: subject noun + verb + numeral + classifier + object noun. They manipulated the semantic (in)congruency in the classifier-noun, verb-noun, and verb-classifier at different levels of syntactic hierarchy during sentence processing. The results revealed significant N400 effects for nouns in all three mismatch conditions, indicating semantic processes at both lower- and higher-level syntactic structures. However, Chan (2019) challenged the interpretation of the N400 effect observed in the classifier-noun mismatch reported by Zhou et al. (2010). Chan argued that it might not be a typical N400, but rather a confounding result of LAN (left anterior negativity) and N400 components, “signaling the coexistence of morphosyntactic and semantic agreement in processing classifier phrases” (Chan, 2019, p. 14). This interaction between semantic and syntactic processes in classifier-noun agreement has also been observed in Chou et al. (2015).

Despite the controversial nature of classifier-noun agreement, researchers agree that classifiers semantically constrain their following nouns in features such as animacy, shape, or size, and that comprehenders can pre-activate semantic features of the upcoming nouns when they encounter classifiers (Kwon et al., 2017; Grüter et al., 2019; Huang et al., 2023). For example, Huang et al. (2023) investigated whether comprehenders use classifiers to predict the animacy of upcoming nouns. Native Chinese speakers were presented with animate-constraining (e.g., 位, wei) or inanimate-constraining classifiers (e.g., 本, ben) followed by congruent or incongruent nouns. ERP analyses revealed an N400 effect for incongruent conditions, indicating the difficulty of semantic integration with incompatible nouns, consistent with previous studies (Zhou et al., 2010; Chou et al., 2015; Chan, 2019). Additionally, representational similarity analysis showed that neural activity patterns were more similar following animate-constraining classifiers than inanimate-constraining classifiers before noun presentation, driven by perditions from classifiers about the animacy features of the upcoming nouns.

In addition, Kwon et al. (2017) explored the classifiers' prediction effect in a highly constraining discourse context using relative clauses (e.g., 张艺谋指导的… zhangyimou zhidao de…, “Zhang Yimou directed…”), which set an expectation for a specific head noun. The head nouns in these clauses were manipulated to be either predicted (电影, dianying, “movie”) or non-predicted (e.g.,大厦, dasha, “building”), and paired with appropriate classifiers (部, bu for movie; 座, zuo for building), establishing classifier-noun agreement. The results demonstrated that participants pre-activated semantic features of the predicted head noun (电影, dianying, “movie”) when they encountered semantically constraining relative clauses. This was evidenced by an N400 effect when a mismatched classifier (座, zuo) was read, reflecting challenges in processing a classifier incongruent with the pre-activated semantic information of the predicted noun.

What's intriguing about the classifier-noun agreement is that the classifier and its associated noun can be separated from each other, creating a long-distance dependency. In Mandarin, noun phrases strictly follow a head-final structure, requiring determiner-classifier sequences (such as demonstratives, numerals, and classifiers) and modifiers (adjectives and relative clauses) to precede the head noun. This structural characteristic enables the insertion of a modifying relative clause between the classifier and its associated noun, as demonstrated in (5).

(5) a. Intervening by a Subject-gap RC

   

In (5a), a subject-gap RC is inserted between the classifier tiao “CL(long-shaped objects/animals)” and its associated noun yu “fish”, creating a long-distance dependency between them. Similarly, (5b) establishes a long-distance dependency between the classifier tiao and its associated noun, but also introduces a temporary local incongruity due to the additional noun linjun “neighbor” serving as the embedded subject in the object-gap RC. This issue of syntactic uncertainty and potential local semantic incongruity is also known as the garden path problem for processing head-final relative clauses (Lin and Bever, 2011; Lin, 2019).

Researchers (e.g., Hsu et al., 2014; Wu et al., 2018) have investigated whether the temporary mismatch between the classifier and noun as in (5b) can effectively serve as a cue for predicting an RC structure. For example, Hsu et al. (2014) investigated the impact of temporary classifier-noun incongruity on RC disambiguation by introducing an object-gap RC between a classifier and the head noun in a head-final RC structure. An ERP study revealed that the initial incongruity between the classifier and its adjacent noun did not guarantee an RC prediction. Wu et al. (2018) explored the same issue by manipulating the clause-initial classifier in a self-paced reading task. Their findings demonstrated that encountering temporary incongruity between a classifier wei “CL(people)” and its adjacent noun shitou “stone” in a sentence-initial position would lead the reader to postulate a phrasal boundary between them and anticipate a suitable noun (in this case, a human) that agrees with the classifier in the upcoming words. Moreover, they found that the temporary incongruity resulted in faster reading times in the region following the RC head noun, indicating that the initial classifier-noun mismatch facilitated the anticipation of an RC structure, albeit the effect was observed in the spill-over region.

Overall, the existing evidence indicates that classifiers can serve as cues for predicting the semantic features of upcoming nouns on a phrasal or sentential level (e.g., relative clauses). In other words, people can pre-activate semantic features such as animacy, humanness, size, or shape of the upcoming nouns when they encounter classifiers before the nouns appear, due to the semantic and syntactic dependencies between classifiers and their associated nouns in Mandarin Chinese. However, the extent to which the dependency between classifiers and the upcoming nouns can modulate RC attachment preferences remains underexplored. Therefore, this study investigated whether classifier-noun agreement can function as a morphosyntactic predictor of semantic dependencies and influence relative clause attachment preferences.

1.4 The present study

This study conducted two self-paced reading experiments to investigate the influence of an initial classifier on comprehenders' anticipation of its associated noun and the impact of this prediction on relative clause attachment preferences in Mandarin Chinese. This study involved the insertion of a subject-gap RC between a demonstrative-classifier phrase (DEM-CL) and a complex RC head noun (NP_low de NP_high, “NP_high of NP_low”), resulting in the sequence “DEM-CL [gap VP de1]_RC NP_low de2 NP_high.” For instance, in sentence (6), a subject-gap RC (who is drinking water) was inserted between the demonstrative-classifier phrase (zhe-ming “this-CL(people)”) and a complex nominal phrase (yigong de xiaomao “volunteer's cat”).

(6) Zheming    zhengzai   de   yigong   de   xiaomao

            heshui

  DEM-      RC        de1  NP_low   de2  NP_high

  CL(people)   

  This-     is       relati-  volunteer poss.   cat

  CL(people)    drinking  vizer   

      water

‘the cat of the volunteer who is drinking water.'   

Note that the two occurrences of the functional morpheme DE realized as de1 and de2 are distinctive in their grammatical functions and interpretations, as de1 is the relativizer of the RC while de2 is a prenominal modifier indicating a possessive (genitive) relationship between NP_low and NP_high. Crucially, the DEM-CL phrase is placed before the relative clause to generate a long-distance classifier-noun agreement. In addition, since previous studies have shown that animacy interacts with RC types and can affect RC attachment (Kwon et al., 2019), in this experiment, we controlled for both animacy and RC type by adopting only animate nouns to serve the head nouns in subject-gap RCs. We also set up NP_low to be semantically [+human] and NP_high to be [+animal], and there exist the same possessive relationships between NP_low and NP_high in all experimental items.

Since attachment preferences are ultimately related to temporarily storing RCs and NPs in the working memory for interpretation, the present study also explores the potential role of working memory capacity (WMC) as an individual difference variable. The impact of WMC on RC attachment preferences has long been studied, but conflicting results still exist. Some offline studies have shown a link between low WMC and a preference for high attachment. For instance, Mendelsohn and Pearlmutter (1999) found that low-span subjects in English preferred high attachment, a finding supported by subsequent research (Swets et al., 2007; Cotter and Ferreira, 2014). However, other offline studies (German: Harding et al., 2019; English and Persian: Mahmoodi et al., 2022) do not find a correlation between WMC and RC attachment preferences. Similarly, in online tasks, some studies have shown that low-span subjects are inclined to attach RCs to the high-attachment site, compared to high-span subjects (English: Marefat et al., 2015). However, other studies have reported the opposite pattern (Traxler, 2007), and some have found no effect of WMC (Omaki, 2005; Traxler, 2009). Therefore, the present study incorporates WMC, measured by an adapted reading-span task for Mandarin sentence processing (Daneman and Carpenter, 1980) to investigate its impact in moderating RC attachment preferences in both offline and online stages.

2 Experiment 1

Experiment 1 looked at comprehenders' RC attachment preferences in Mandarin Chinese during both online and offline processing. We also studied how the presence of an initial classifier pre-activate semantic information of its associated nouns, thus impacting comprehenders' RC attachment preferences.

In this experiment, we manipulated the agreement between the classifier and the two potential attachment sites, NP_low and NP_high, and created four conditions: (i) CL-NP_low, where the classifier only agrees with NP_low, enforcing a low attachment; (ii) CL-NP_high, where the classifier only agrees with NP_high, enforcing a high attachment; (iii) CL-NP_low/NP_high, where the classifier could agree with either NP_low or NP_high; and (iv) No CL, representing a condition without a classifier. Both the CL-NP_low/NP_high and the No CL condition create global ambiguity in terms of RC attachment preferences. Table 1 demonstrates the four conditions with an example. See also Figure 3 for illustrations on syntactic structures regarding the exemplar item. In the CL-NP_low condition, the classifier ming “CL(people)” agrees only with the first noun NP_low yigong “volunteer” but not with NP_high xiaomao “cat”. In the CL- NP_high condition, the classifier zhi “CL(animal)” agrees solely with the second noun NP_high xiaomao “cat” but not with NP_low yigong “volunteer”. The CL-NP_low/NP_high condition introduced ambiguity by allowing the group classifier qun “CL(group)” to be congruent with either NP_low or NP_high. Finally, the No CL condition does not involve a classifier.

Table 1

Table 1. An experimental item in the four conditions.

Figure 3

Figure 3. Relative clause attachment structures regarding the exemplar item in Mandarin [high-attachment: the upper (A); low- attachment: the lower (B)].

Drawing on the prediction effect in language processing (Kuperberg and Jaeger, 2016; Pickering and Gambi, 2018), and considering Mandarin's SVO word order and head-final RC structure, we predict that comprehenders pre-activate semantic features of the upcoming noun upon encountering the sentence-initial classifier. This anticipation persists until they come across the two potential attachment noun sites. We hypothesize that the classifier-noun agreement, similar to number agreement in English or gender agreement in Spanish, would serve as a disambiguating factor influencing comprehenders' RC attachment preferences. When the classifier holds a semantic dependency with the lower noun NP_low, the RC may also preferably be attached to NP_low. However, when the classifier holds a semantic dependency with the higher noun, NP_high, the RC is more likely to be attached high to this non-local NP. By manipulating the agreement between the initial classifier and either of the two potential attachment sites (NP_low or NP_high), we may observe the semantic influence of classifiers on comprehenders' preferences for high or low RC attachments.

2.1 Methods

2.1.1 Participants

Eighty native Mandarin Chinese speakers (62 females and 18 males) participated in this experiment, recruited via Chinese social media and university classified ads in China. Participants' ages ranged from 18 to 25 years old. Informed consent was obtained for experimentation with participants. Compensation was provided for their participation. The entire study was approved by IRB (##2005886743).

2.1.2 Materials

Twenty-four experimental items were created,¹ all following the same syntactic structure and segmented into eleven regions (see Table 1 for an example). Six critical regions of interest are: de1, NP_low, de2, NP_high, ADV_main and VP_main. All the experimental items were counterbalanced using a Latin-square design. Each participant was presented with one version of each experimental item. Additionally, 56 unambiguous distractors were also included in this experiment. These included 14 subject-gap RCs with animate head nouns, 14 object-gap RCs with inanimate head nouns, 14 non-RC sentences featuring a passive BEI construction (“be-passive”), and 14 non-RC sentences with an existential YOU construction (“there exits…”). Each participant encountered a total of 80 trials, presented in a random order.

To enhance the natural flow of the stimuli and to prime participants' awareness of classifier-noun agreement prior to introducing each test item, a referential discourse context was provided. This context included equal mentions of classifier nominal phrases for both NP_low and NP_high twice. For instance, the context for the exemplar item in Table 1 would be “A CL(people) volunteer adopted a CL(animal) stray cat. One day, the CL(people) volunteer brought this CL(animal) cat to my house to play.” The contexts remained consistent across all four conditions for each item, with the exception of the CL- NP_low/NP_high condition where the classifiers for people and animals were replaced with the classifier qun “CL(group)”.

Before carrying out the experiment, two norming tests were conducted to validate the acceptability of classifier-noun agreement and the equitable use of relative clauses to modify NP_low or NP_high. Twenty-five native Mandarin Chinese speakers were paid to evaluate all the classifier-noun pairs used in both Experiment 1 and Experiment 2. Participants were required to rate the acceptability of classifier-noun pairs on a scale ranging from 1 to 7, with 1 signifying complete unacceptability and 7 indicating total acceptability. Only classifier-noun pairs rated above 6 were incorporated into the experiments: CL- NP_low (people) (M = 6.71; SD = 0.19); CL- NP_high (animal) (M = 6.90; SD = 0.15); Qun- NP_low (people) (M = 6.65; SD = 0.17); Qun- NP_high (animal) (M = 6.73; SD = 0.16).

In addition, two surveys were created using a Latin-square design to assess the modifying plausibility of relative clauses. Each survey contained 24 items from Experiment 1 and 56 items from Experiment 2 (will be discussed further in the Experiment 2 section). To maintain a balance between acceptable and unacceptable items, 40 fillers of comparable length and structure were included. These fillers were intentionally crafted to be semantically unacceptable, for example, “the goldfish that is watching TV on the couch.” A distinct group of 20 native speakers was randomly assigned to one of the surveys. Each participant assessed the acceptability of relative clauses modifying either [+human] nouns (e.g., “the volunteer that is drinking water on the balcony”) or [+animal] nouns (e.g., “the cat that is drinking water on the balcony”) on a 7-point Likert scale, where 1 indicated complete unacceptability and 7 represented full acceptability. The results demonstrated that relative clauses in Experiment 1 could modify both NP_low (people) (M = 6.65; SD = 0.33) and NP_high (animal) (M = 6.58; SD = 0.40) with equal acceptability. A two-tailed paired t-test revealed no significant differences between them (t (23) = 0.16, p = 0.87).

2.1.3 Procedure

In this experiment, a moving-window self-paced reading task was developed using jsPsych Version 6.2 (de Leeuw et al., 2023), and executed on the Cognition.run platform. Participants were presented with a context in the center of the screen. After reading the context, they could proceed by pressing the spacebar. The context would then be replaced by a fixation “+”. Pressing the spacebar again allowed participants to advance to the next page, where the fixation was replaced by a dashed line. Participants pressed the spacebar to move through the successive displays of the target sentence, with their reaction times recorded for each region.

Following the target sentence, participants advanced to the comprehension question phase, where they were presented with a forced-choice question regarding the RC attachment. For instance, “____is drinking water on the balcony? F: volunteer; J: cat”. The relative order of the two options was balanced across test trials, with half of them having F taken by NP_low and J taken by NP_high, and the other half with the reverse order. Participants' choices were recorded during this phase. To familiarize participants with the task procedure, three practice trials were administered before the actual experiment.

After completing the self-paced reading task, participants proceeded to a reading span task designed to assess their working memory capacity, adapted from Daneman and Carpenter (1980). This task comprised five sets of sentences, with three trials in each set. The number of sentences in each set ranged from 2 to 6, representing different reading span sizes. Thus, a total of 60 sentences (2 sentences per trial ^* 3trials + 3sentences per trial ^*3 trials + 4 sentences per trial ^*3 trials + 5 sentences per trial ^*3 trials + 6 sentences per trial ^*3 trials) was included in the reading span task. Participants saw one sentence at a time on the screen and read it aloud. After reading all the sentences in a trial, they were required to recall the last word of each sentence in the correct order. If they couldn't remember a word, they were instructed to type “do not remember.” Only words typed correctly in both form and order were considered accurate and recorded. The total number of words recalled correctly was calculated for each participant. Following the reading span task, participants were directed to complete a background questionnaire.

2.1.4 Data analysis

The data obtained from the self-paced reading task and the reading span task underwent preprocessing before any statistical analysis. Initially, the reading span scores were standardized using z-scores. Regarding comprehension question choices, NP_low (indicating low attachment) was coded as 0, while NP_high (indicating high attachment) was coded as 1. In the data analysis phase, a generalized linear mixed-effects model was applied to investigate participants' comprehension question responses, with “condition” (reflecting classifier-noun agreement: CL-NP_low; CL-NP_high; CL-NP_low/NP_high; No CL) and “reading span z-score” (indicating working memory capacity) as fixed factors. Furthermore, both “subject” and “item” were considered as random factors. Since “condition” was a four-level categorical predictor, we converted it into three contrasts. The first contrast compares “No CL” with “CL-NP_low” (No CL = −0.5, CL-NP_low = 0.5, CL-NP_high = 0, CL-NP_low/NP_high = 0). The second contrast compares “No CL” with “CL-NP_high” (No CL = −0.5, CL-NP_high = 0.5, CL-NP_low = 0, CL-NP_low/NP_high = 0). The third contrast compares “No CL” with “CL-NP_low/NP_high” (No CL = −0.5, CL-NP_low/NP_high = 0.5, CL-NP_low = 0, CL-NP_high = 0). We first created the most complex model with the maximal random-effects structure, justified by the design (Barr et al., 2013), as shown in (7).

(7) Complex model = glmer [cq_choice ~ condition+

 rs_zscore + (1 + condition|subject) + (1 + condition +

 rs_zscore|item), data = data, family = “binomial”

 (link =“logit”)]

Convergence failures or singularity warnings were dealt with by progressively simplifying the random effects structure until convergence was reached or singularity warning disappeared (Barr et al., 2013; Linck and Cunnings, 2015). For model selection, we employed a stepwise backward elimination method using the Akaike Information Criterion (AIC, Akaike, 1998). We began with the most complex model and reduced the model complexity until no further improvement in model fit is observed (Matuschek et al., 2017). Considering the design of Experiment 1, the variables of interest were “condition” and “reading span z-score.” These variables were included in the models as fixed factors and were retained throughout the model selection procedure. In addition, post hoc tests were performed with Bonferroni correction for any significant interactions that were identified. For details on the model selections and AIC values, please refer to the R script (find the link under the Data Availability section).

In preprocessing the reading times, we first eliminated outliers using absolute cutoffs, excluding reading times < 100 ms or >3,000 ms (Luce, 1986; Roberts and Felser, 2011), constituting 6.13% of the total data. Subsequently, we log-transformed the reading times, a standard practice to stabilize variance and achieve approximately normal distribution of residuals (Box and Cox, 1964). Linear mixed-effects models were then conducted on the log-transformed reading times in each region, considering “condition” and “reading span z-score” as fixed factors and “subject” and “item” as random factors. The same contrast-coding for “condition” was applied here. We first created the most complex model with the maximal random-effects structure, justified by the design (Barr et al., 2013), as shown in (8).

(8) Complex model = lmer(logRT ~ condition + rs_zscore +

 (1+condition|subject) + (1+condition + rs_zscore|item),

 data = data)

Convergence failures, singularity warnings, model selections, and post hoc tests were managed following the same principles as previously outlined. Refer to the R script for model selections and AIC values. We utilized the lme4 package version 1.1-34 and the lmerTest package version 3.1-3 for analyses. All statistical analyses were conducted in R version 4.3.1 (R Core Team, 2023).

2.2 Results and discussion

2.2.1 Offline comprehension questions

The selected model for comprehension question choices is shown in (9), with an AIC value of 1942.7. The by-subject random slope was eliminated due to convergence issues.

(9) Selected model = glmer [cq_choice ~ condition+

 rs_zscore + (1|subject) + (1+ rs_zscore |item), data =

 data, family = “binomial” (link =“logit”)]

The participants' comprehension question choices in each condition are shown in Figure 4, and statistical results are given in Table 2. Figure 4 illustrates a prevailing NP_high (High-Attachment) preference across all conditions: 62.5% in the CL-NP_low condition, 71.9% in the CL-NP_high condition, 66.9% in the CL-NP_low/NP_high condition, and 63.1% in the No CL condition. Table 2 further revealed that the selection of NP_high in the CL-NP_low condition was significantly lower than that in the No CL condition (β = −0.49, SE = 0.21, z = −2.37, p < 0.05, OR² = 0.62, 95% CI³ [−0.89, −0.08]). However, the selection of NP_high in the CL-NP_high condition was significantly higher than that in the No CL condition (β = 0.85, SE = 0.21, z = 3.40, p < 0.001, OR = 2.34, 95% CI [0.43, 1.27]). No significant differences were observed between the CL-NP_low/NP_high condition and the No CL condition (β = 0.09, SE = 0.21, z = 0.41, p = 0.680, OR = 1.09, 95% CI [−0.32, 0.49]). In addition, there was no significant effect of working memory capacity (i.e., rs_zscore) on the comprehension question responses (β = 0.19, SE = 0.23, z = 0.81, p = 0.416, OR = 1.21, 95% CI [−0.26, 0.64]).

Figure 4

Figure 4. Percentage of comprehension question choices by conditions.

Table 2

Table 2. Generalized linear mixed-effects model for comprehension question responses.

To better understand the impact of “condition” on participants' comprehension responses, the data was also fitted with a model using treatment coding for the categorical predictor “condition”. Pairwise comparisons conducted via the emmeans package (version 1.10.2) revealed a significant difference between the CL-NP_low and CL-NP_high conditions. The selection of NP_high was significantly higher in the CL-NP_high condition than in the CL-NP_low condition (β = −0.67, SE = 0.17, z = −3.89, p < 0.001, OR = 0.51, 95% CI [−1.01, −0.33]).

2.2.2 Online reading times

Table 3 presents the means and standard deviations of raw reading times across four conditions in six regions of interest. Figure 5 displays a line graph depicting log-transformed reading times. Linear mixed-effects models were applied to the six regions of interest. The results showed no significant effects of “condition” or “reading span z-score” in the de1, NP_low, and NP_high regions. However, significant differences were observed between the “No CL” condition and the “CL-NP_high” condition in the spill-over region of NP_low (i.e., de2), as well as in the two spill-over regions of NP_high (i.e., ADV_main and VP_main).

Table 3

Table 3. The mean reading times and standard deviations in the six regions of interest (milliseconds).

Figure 5

Figure 5. Line graph of reading times in Experiment 1 (with 95% confidence intervals).

Table 4 presents the statistical results with the selected models using contrast-coding for each region. As Table 4 demonstrates, in the de2 region, reading times in the CL-NP_high condition were notably longer than those in the No CL condition (β = 0.03, SE = 0.01, t = 2.20, p < 0.05, d = 0.174,⁴ 95% CI [0.002, 0.05]). In contrast, reading times in the CL-NP_high condition were significantly shorter than those in the No CL condition in the ADV_main region (β = −0.03, SE = 0.02, t = −2.08, p < 0.05, d =-0.17, 95% CI [−0.06, −0.002]), and the VP_main region (β = −0.06, SE = 0.01, t = −4.09, p < 0.001, d = −0.33, 95% CI [−0.09, −0.03]).

Table 4

Table 4. Linear mixed-effects models in de2, ADVmain, and VPmain regions.

In addition, reading times in the six regions were fitted to models with treatment coding for the categorical predictor “condition” to explore pairwise comparisons. No significant differences emerged in the de1, NP_low, and NP_high regions. In the de2 region, reading times in the CL-NP_high condition were significantly longer than those in the CL-NP_low/NP_high condition (β = 0.02, SE = 0.01, t = 2.42, p < 0.05, d = 0.16, 95% CI [0.005, 0.04]). In the ADV_main region, reading times in the CL–NP_high were significantly shorter than in the CL–NP_low condition (β = −0.03, SE = 0.01, t = −2.13, p < 0.05, d = −0.14, 95% CI [−0.05, −0.002]). Similarly, in the VP_main region, reading times in the CL–NP_high were significantly shorter than in the CL–NP_low condition (β = −0.04, SE = 0.01, t = −3.23, p < 0.01, d = −0.21, 95% CI [−0.06, −0.02]) and the CL–NP_low/NP_high condition (β = −0.04, SE = 0.01, t = −3.51, p < 0.01, d = −0.23, 95% CI [−0.07, −0.02]). No other significant differences were found across levels in these regions. Additionally, no significant effect of “reading time z-score” was observed in any of the six regions.

In this experiment, we explored people's preferences for RC attachment ambiguity, examining both online and offline processing phases. Additionally, we investigated whether an initial classifier pre-activates the semantic features of the upcoming noun, thereby influencing RC attachment preferences. To do this, we manipulated the agreement between the classifier and potential attachment sites (NP_low and NP_high), creating four conditions: CL-NP_low, CL-NP_high, CL-NP_low/NP_high, and No CL.

The offline comprehension results indicated a prevalent preference for high attachment across all four conditions, revealing an intrinsic inclination toward high attachment during the offline processing phase. However, significant differences emerged: the selection of NP_high in the CL-NP_high condition was markedly higher than in the No CL condition, whereas the selection of NP_high in the CL-NP_low condition was significantly lower than in the No CL condition. These findings highlight the influence of classifier-noun agreement on people's RC attachment choices. The CL-NP_high agreement increased the intrinsic high-attachment preference, while the CL-NP_low agreement decreased it. Consequently, a significant difference was observed between the CL-NP_low and CL-NP_high conditions in terms of comprehension question choices. Additionally, no significant difference was observed between the CL-NP_low/NP_high condition and the No CL condition, further confirming that the group classifier qun “CL(group)” in the CL-NP_low/NP_high condition was ambiguous, functioning similarly to the No CL condition.

These findings were further confirmed in the online processing phase. Before delving into the online data, it's worth noting again that the word order of relative clauses in Mandarin is reversed compared to English. In Mandarin, the RC precedes the complex head noun, and the low-attachment site (NP_low) precedes the high-attachment site (NP_high). These features of Chinese RC structure suggest that when encountering NP_low first, the competitor (NP_high) has not yet appeared. Therefore, it would be inappropriate to describe low attachment as being “preferred” at this point in an online study of parsing ambiguity since there is no attachment conflict at this early stage. Therefore, we followed Kamide and Mitchell (1997, 1999) in attributing this initial “low-attachment” to the incremental processing effect (Kamide and Mitchell, 1997, 1999). However, we do use the term “preference” for the high-attachment site if it turns out to attract attachment, as both attachment sites are now present, and readers can demonstrate their preferences either by maintaining the initial low-attachment or shifting to high-attachment.

Analyses of the self-paced reading times revealed that readers pre-activate semantic features of the upcoming nouns when they encounter the classifiers. These predictions are maintained until they encounter the two potential attachment nouns, affecting how people process these sites and modulating their RC attachment preferences.

Specifically, in the spill-over region of NP_low (i.e., de2), reading times in the CL-NP_high condition were significantly longer than in the No CL condition. In contrast, in the spill-over regions of NP_high (i.e., ADV_main and VP_main), reading times in the CL-NP_high condition were significantly shorter than in the No CL condition. These results indicate that when the sentence-initial classifier is CL(animal) (e.g., 只, zhi), it pre-activates the semantic features of an animal for the upcoming noun, and this prediction persists until comprehenders encounter the first potential noun, NP_low(people) (e.g., 义工, volunteer), which conflicts with the anticipation and thus creates processing difficulty. This is evidenced by the significantly increased reading times in the spill-over region compared with the control condition. The incongruence between the CL(animal) (e.g., 只, zhi) and NP_low(people) (e.g., 义工, volunteer) causes comprehenders' expectation for an animal noun to persist until encountering the second potential noun, NP_high(animal) (e.g., 小猫, cat). The pre-activated semantic features of an animal match with NP_high(animal) (e.g., 小猫, cat), and the congruence between CL(animal) and NP_high(animal) facilitates processing, manifesting as significantly decreased reading times in the spill-over regions compared to the control condition.

On the other hand, when the classifier is CL(people) (e.g., 位, wei), it pre-activates the semantic features of people for the upcoming noun, persisting until the first potential noun NP_low(people) (e.g., 义工, volunteer), which matches well with the classifier's anticipation. As a result, the NP_low(people) (e.g., 义工, volunteer) integrates smoothly with the preceding sentence, and no processing difficulty emerge in the NP_low and its spill-over de2 regions. After this integration, when comprehenders encounter an additional noun, NP_high(animal) (e.g., 小猫, cat), this additional noun creates integration difficulty, resulting in increased reading times. Therefore, the sentence-initial classifiers provide semantically constraining contexts for readers to make predictions about the upcoming noun, thus creating forced-low-attachment or forced-high-attachment conditions for RC interpretation.

In addition, the globally ambiguous conditions, namely the “No CL” and “CL-NP_low/NP_high” conditions, allow us to observe the locus of people's intrinsic attachment preferences from the online phase to the offline phase. In these two conditions, self-paced reading times in the NP_low region indicate an initial low-attachment during early online processing due to “incremental pre-head attachment parsing” (Kamide and Mitchell, 1999, p. 635). Offline comprehension questions, however, show a high-attachment preference. This inconsistency between the initial online low-attachment and the eventual offline high-attachment preference suggests that the initial low attachment is reconsidered during later online processing, influenced by discourse or post-syntactic information available later in the sentence, consistent with previous studies (de Vincenzi and Job, 1993; Kamide and Mitchell, 1997, 1999; Fernández, 2002).

This dynamic preference pattern has also been confirmed by the two disambiguated conditions, i.e., CL-NP_low and CL-NP_high conditions. According to Fernández (2002), readers' attachment preferences in the online processing phases would manifest as increased reading times at critical regions when encountering materials that contain the dis-preferred forced attachment, as compared to reading times associated with either ambiguous materials or those containing the preferred forced attachment (Fernández, 2002). In this experiment, we observed that in the critical region, NP_high and its spill-over regions (ADV_main and VP_main), reading times in the forced-high-attachment condition (i.e., CL-NP_high) were significantly shorter than in the forced-low-attachment condition (CL-NP_low) and the ambiguous condition (No CL), indicating a high-attachment preference in the later online processing stage. Therefore, readers initially show low-attachment, which then switches to a high-attachment preference in the later stage of online processing due to discourse or post-syntactic information. This high-attachment preference remains until the offline phase.

3 Experiment 2

In Experiment 2, we further investigated the prediction effects of classifiers in resolving RC attachment ambiguity, particularly when an additional disambiguating cue, namely semantic compatibility, was introduced. This is because the RCs can be attached to both the low and the high NPs in Experiment 1, thus making it challenging to ensure how participants interpret these RCs in addition to relying on the semantic cues on the classifiers. Our aim was to discern the relative significance of each cue in predicting the upcoming noun and shaping attachment preferences.

Following the syntactic structure of Experiment 1, we introduced semantic compatibility as an additional variable alongside classifier-noun agreement in Experiment 2. Semantic compatibility was manipulated between the RC and its potential attachment sites, where each RC was exclusively compatible with either NP_low or NP_high. The variable of “semantic compatibility” therefore has two levels: that the RC is semantically compatible with NP_low (RC-NP_low) or that the RC is semantically compatible with NP_high (RC-NP_high). We also manipulated the classifier-noun agreement, including CL-NP_low, CL-NP_high, CL-NP_low/NP_high, and No CL levels. Thus, Experiment 2 had a 2^*4 design incorporating two independent variables: semantic compatibility (2 levels) and classifier-noun agreement (4 levels), resulting in eight experimental conditions initially. However, due to the issue of crossing dependency, one of the conditions (RC-NP_high & CL-NP_low) was not possible and was eliminated. To maintain a factorial design, we therefore excluded the level of CL-NP_low. Consequently, Experiment 2 adopted a 2^*3 design with two independent variables: semantic compatibility (RC-NP_low and RC-NP_high) and classifier-noun agreement (CL-NP_high, CL-NP_low/NP_high, and No CL), establishing six experimental conditions. Table 5 illustrates an exemplar item in these six conditions.

Table 5

Table 5. An experimental item in the six conditions.

We predict that when the RC aligns semantically with NP_high (i.e., RC-NP_high) and the classifier corresponds with NP_high (i.e., CL-NP_high), the coherence between semantic compatibility and classifier-noun agreement promotes a strong prediction for NP_high, resulting in a high-attachment preference. However, complications arise when the RC is semantically compatible with NP_low (i.e., RC-NP_low), but the classifier aligns with NP_high (i.e., CL-NP_high), generating a conflict in predictions of the upcoming noun. In such scenarios, the dominance of one cue prevails, establishing either a low-attachment preference or a high-attachment preference. Finally, in cases where the classifier is ambiguous (CL-NP_low/NP_high) or absent (No CL), the RC attachment preference is substantially predicted by the semantic compatibility between the RC and the potential noun phrases.

3.1 Methods

3.1.1 Participants

Sixty native Mandarin Chinese speakers aged between 18 and 30 participated in this experiment. Participants received compensation for their involvement, and informed consent was obtained from all participants.

3.1.2 Materials

Twenty-eight experimental items were created, all following the same syntactic structure and segmented into eleven regions, as illustrated in Table 6. Among these regions, six critical regions of interest were identified: de1, NP_low, de2, NP_high, ADV_main, and VP_main. All the experimental items were counterbalanced using a Latin-square design. Each participant was presented with one version of each experimental item. The experiment incorporated the same set of 56 fillers utilized in Experiment 1.

Table 6

Table 6. The average reading times and standard deviations in the six regions of interest (milliseconds).

In line with the paradigm in experiment 1, each target sentence was introduced by a referential discourse context. To ensure the exclusive semantic compatibility of RCs with either NP_low or NP_high, the materials used in this experiment underwent validation using the same norming tests employed in Experiment 1. Participants utilized a 7-point Likert scale, ranging from 1 denoting complete unacceptability to 7 indicating full acceptability. The results of these acceptability judgments revealed a clear pattern: half of the RCs demonstrated semantic compatibility with NP_high (M = 6.78; SD = 0.24) but not NP_low (M = 1.20; SD = 0.21) [t₍₂₇₎ = 98.09, p < 0.001]. Conversely, the other half of the RCs exhibited compatibility with NP_low (M = 6.80; SD = 0.23) but not NP_high (M = 1.27; SD = 0.25) [t₍₂₇₎ =88.09, p < 0.001].

3.1.3 Procedure and data analysis

In this experiment, we employed a moving-window self-paced reading task created using jsPsych Version 6.2 (de Leeuw et al., 2023) and executed on the Cognition.run platform. The experimental procedure followed that of Experiment 1. Prior to any statistical analysis, the data collected from both the self-paced reading task and the reading span task underwent preprocessing. Initially, reading span scores were standardized into z-scores. The accuracy rate for comprehension questions averaged 97.91% (SD = 2.68). Outliers were then removed using absolute cutoffs, excluding reading times < 100 ms or >3,000 ms (Luce, 1986; Roberts and Felser, 2011), accounting for 2.61% of the total data. Subsequently, reading times were log-transformed. Linear mixed-effects models were applied to log-transformed reading times, incorporating “semantic compatibility” and “CL-N agreement” as interacting predictors, with “reading span z-score” as a co-predictor. Since “CL-N agreement” was a three-level categorical predictor, we converted it into two contrasts: the first compared “No CL” with “CL-NP_high” (No CL = −0.5, CL-NP_high = 0.5, CL-NP_low/NP_high = 0), and the second compared “No CL” with “CL-NP_low/NP_high” (No CL = −0.5, CL-NP_low/NP_high = 0.5, CL-NP_high = 0). In addition, “semantic compatibility” was a two-level categorical predictor, and it was contrast-coded (RC-NP_low = 0.5, RC-NP_high = −0.5). We first created the most complex model with the maximal random-effects structure, justified by the design (Barr et al., 2013), as shown in (10).

(10) Complex model = lmer [logRT ~ cl_n_agreement ^*

 rc_semantics + rs_zscore + (1+ cl_n_agreement ^*

 rc_semantics |subject) + (1+ cl_n_agreement ^*

 rc_semantics + rs_zscore|item), data = data]

Convergence failures, model selections, and post hoc tests were managed following the same principles as in Experiment 1. Considering the design of Experiment 2, the variables of interest were “classifier-noun agreement”, “semantic compatibility”, and “reading span z-score.” These variables were included in the models as fixed factors and were retained throughout the model selection procedure.

3.2 Results and discussion

Table 6 displays the means and standard deviations of raw reading times across six conditions in six regions of interest. Meanwhile, Figure 6 presents a line graph illustrating log-transformed reading times. Linear mixed-effects models were fitted in the six regions of interest.

Figure 6

Figure 6. Line graph of reading times in Experiment 2 (with 95% confidence intervals).

The results showed no significant effects of “classifier-noun agreement”, “semantic compatibility”, or “reading span z-score”, nor interaction effects in the de1 and NP_high regions (ps > 0.05). However, a significant effect of “semantic compatibility” was observed in the NP_low region, its spill-over de2 region, as well as in the spill-over region of NP_high (VP_main). Table 7 presents the statistical results from selected models using the contrast-coding approach (also see Figure 7 for illustration).

Table 7

Table 7. Linear mixed-effects models in NP_low, de2, and VP_main regions.

Figure 7

Figure 7. Line graph of reading times with 95% confidence intervals illustrating the semantic compatibility effect.

In the NP_low region, reading times in the RC-NP_high condition were significantly longer than those in the RC-NP_low condition (β = −0.02, SE = 0.01, t = −2.81, p < 0.01, d = −0.15, 95% CI [−0.04, −0.01]). No significant effects of “classifier-noun agreement” or “reading span z-score” were observed, and there was no interaction effect in this region.

Moving to the spill-over de2 region, reading times in the RC-NP_high condition were significantly longer than those in the RC-NP_low condition (β = −0.03, SE = 0.01, t = −3.60, p < 0.001, d = −0.20, 95% CI [−0.05, −0.01]). In addition, an interaction effect between “classifier-noun agreement” and “semantic compatibility” was observed in the de2 region. To further explore the interaction effect, the data was fitted with a model using treatment coding for the two categorical predictors, and pairwise comparisons between conditions were conducted. Results showed that reading times in the “RC-NP_high CL-NP_high” condition were significantly longer than those in the “RC-NP_low No CL” condition (β = −0.05, SE = 0.01, t = −3.72, p < 0.01, d = −0.33, 95% CI [−0.08, −0.02]), the “RC–NP_low CL–NP_low/NP_high” condition (β = −0.05, SE = 0.01, t = −3.45, p < 0.01, d = −0.31, 95% CI [−0.07, −0.02]), the “RC–NP_low CL–NP_high” condition (β = −0.06, SE = 0.01, t = −4.29, p < 0.001, d = −0.38, 95% CI [−0.08, −0.03]), and the “RC–NP_high No CL” condition (β = 0.04, SE = 0.01, t = 3.16, p < 0.05, d = 0.27, 95% CI [0.02, 0.06]). Moreover, a marginally significant effect of “reading span z–score” was observed in the de2 region (β = −0.03, SE = 0.01, t = −1.88, p = 0.065, d = −0.18, 95% CI [−0.05, 0.001]).

In the first spill-over region of NP_high, the ADV_main region showed a marginally significant effect of “semantic compatibility.” Reading times in the RC–NP_high condition were shorter than those in the RC–NP_low condition (β = 0.02, SE = 0.01, t = 1.71, p = 0.088, d = 0.09, 95% CI [−0.002, 0.03]). This trend became more pronounced in the second spill–over region of NP_high, the VP_main region, where a significant main effect of “semantic compatibility” was observed. Reading times in the RC–NP_high condition were significantly shorter than those in the RC–NP_low condition (β = 0.02, SE = 0.01, t = 2.23, p < 0.05, d = 0.12, 95% CI [0.002, 0.04]). No other effects were identified in this region.

In this experiment, we examined the impact of two disambiguating cues, namely classifier-noun agreement and the semantic compatibility between the RC and the head noun, on influencing RC attachment preferences. We manipulated the relationship between the prenominal classifier and two potential head nouns, as well as the relationship between RC semantics and the potential head nouns.

Our findings revealed a prediction effect of RC semantics for upcoming nouns, affecting how people process the two potential attachment sites and thus modulating RC attachment preferences. Specifically, a significant main effect of semantic compatibility was observed in the NP_low region, the spill-over region of NP_low (i.e., de2), and the spill-over regions of NP_high (i.e., ADV_main and VP_main). In the NP_low and de2 regions, reading times in the RC-NP_high condition were significantly longer than in the RC-NP_low condition. Conversely, in the ADV_main and VP_main regions, reading times in the RC-NP_high condition were significantly shorter than in the RC-NP_low condition.

These results indicate that the semantics embedded in relative clauses provides a constraining context that predicts a specific upcoming noun. Despite the classifier-noun agreement, when the RC semantics is compatible with NP_high, which is an animal noun, it pre-activates the semantic features of an animal for the upcoming noun. This prediction clashes with the semantics of the first potential noun, NP_low (people), creating processing difficulty and resulting in slower reading times in the NP_low and de2 regions. This incongruence maintains the anticipation for an animal noun until the second potential noun, NP_high (animal) appears. The pre-activated semantic expectation of an animal then matches with the NP_high (animal), facilitating its integration into the preceding sentence and manifesting as fast reading times in the spill-over regions of NP_high.

On the other hand, despite the classifier-noun agreement, when the RC semantics is compatible with NP_low, it leads to the anticipation of a human noun, and this prediction matches with the first potential noun NP_low (people), facilitating the integration of NP_low (people) with the preceding RC, which is evidenced by fast reading times in the NP_low and de2 regions. When readers encounter an additional noun, NP_high (animal), it creates integration difficulty, resulting in slow reading times in the spill-over regions of NP_high.

Furthermore, we observed an interaction effect between classifier-noun agreement and semantic compatibility in the de2 region. Reading times in the condition where both RC semantics and the classifier agreed with NP_high (i.e., “RC-NP_high CL-NP_high”) were significantly longer than those in the condition where RC semantics were compatible with NP_high but there was no sentence-initial classifier (i.e., “RC-NP_high No CL”). This suggests that both RC semantics and sentence-initial classifiers contribute to predictions for the upcoming noun. The prediction effect is notably stronger when both cues contradict with the semantics of the closest NP, namely NP_low, compared to when only RC semantics makes the prediction, underscoring the importance of classifiers' prediction effects. Moreover, reading times in the condition where both RC semantics and the classifier agreed with NP_high (i.e., “RC-NP_high CL-NP_high”) were significantly longer than in the condition where RC semantics were compatible with NP_low but the classifier agreed with NP_high (i.e., “RC-NP_low CL-NP_high”), suggesting the significance of RC semantics prediction effects.

These results indicate that when both cues were present, RC semantic compatibility took precedence over classifier-noun agreement in resolving RC attachment ambiguities, as the classifier effect was only observed in the de2 region, whereas the RC semantic effect was observed in NP_low, NP_low's spill-over region de2, and NP_high's spill-over regions, ADV_main and VP_main. Essentially, RC semantic compatibility emerged as the predominant predicting factor in disambiguating RC attachment resolutions. However, this does not necessarily mean that the classifiers' prediction effect observed in Experiment 1 disappeared. An interaction effect between semantic compatibility and classifier-noun agreement was observed in the de2 region, suggesting that the classifiers' prediction effect did not disappear but instead diminished.

4 General discussion

4.1 Classifier-noun agreement and its prediction effect

The first research question investigates whether classifier-noun agreement can function as a disambiguating cue in resolving relative clauses that contain a head noun embedded within a complex nominal phrase consisting of two noun phrases.

The study found that classifier-noun agreement in Mandarin can disambiguate relative clauses to create forced-high/low-attachment conditions (Fernández, 2002). This finding aligns with previous studies that utilized morphosyntactic dependencies to resolve relative clause attachment ambiguities. For example, Fernández (2002) used number agreement in English, de Vincenzi and Job (1993) used gender agreement in Italian, and Lee and Kweon (2004) used honorific agreement in Korean.

We found that readers use classifier-noun agreement to resolve ambiguous RC attachments. The predictive role of classifiers influences how people anticipate and integrate upcoming information. Specifically, when a prenominal classifier is encountered, it pre-activates specific semantic features of the upcoming noun. For example, if the classifier is CL(animal) (e.g., 只, zhi), it primes the comprehender to expect an animal noun. This pre-activation persists until the expected noun is encountered, guiding the comprehender's expectations and influencing processing efficiency. When the pre-activated semantic features triggered by the classifier and the subsequent noun are congruent (e.g., a classifier predicting an animal followed by a noun denoting an animal), processing becomes faster. This congruence reduces cognitive load, as evidenced by shorter reading times and easier integration into the preceding sentence.

Conversely, when the pre-activated semantic features and the subsequent noun are incongruent (e.g., a classifier predicting animals followed by a noun denoting a person), this mismatch creates processing difficulty. Readers experience increased reading times and cognitive load as they work to reconcile the unexpected noun with their initial predictions. Furthermore, the expectation triggered by the classifier can persist beyond the immediate noun. If the first encountered noun does not match the classifier, the anticipation continues until a matching noun is found.

The observed classifier prediction effect in this study aligns with predictive processing in sentence comprehension, wherein the linguistic features of upcoming words are pre-activated before the bottom-up input of the actual word (Kuperberg and Jaeger, 2016; Pickering and Gambi, 2018). Comprehenders utilize sentence-initial classifiers to provide contextual information, making predictions by pre-activating the semantic features of the upcoming noun before it is encountered (Kwon et al., 2017; Grüter et al., 2019; Chow and Chen, 2020; Huang et al., 2023). This pre-activation enables the brain to anticipate forthcoming linguistic input, promoting faster and more efficient processing (Kuperberg and Jaeger, 2016; Pickering and Gambi, 2018).

In addition, the prediction triggered by classifiers persists until the expected noun is encountered, even if the expected noun is distant from the classifier due to an inserted RC. Upon encountering the expected noun, the classifier's prediction effect facilitates its integration, though the facilitative effect is delayed and observed in the spill-over region of the critical noun. This phenomenon was also noted in Wu et al. (2018).

However, the classifier-noun agreement effect diminished when it co-occurred with a strong disambiguating cue, RC semantics. People primarily rely on semantic information embedded in relative clauses to anticipate and ease the integration of upcoming nouns. In Experiment 2, both the sentence-initial classifier and the relative clause provided disambiguating information, with both preceding the two potential noun sites. This scenario led to both the classifier and the RC making predictions about the upcoming noun. Results indicated that RC semantics played a dominant role in predicting the upcoming noun, as it provided a highly-constraining context that enabled people to make fine-grained predictions, consistent with previous findings (Federmeier and Kutas, 1999; Kamide et al., 2003; Wang et al., 2018).

While RC semantics plays a dominant role in resolving RC attachment ambiguities, the classifier's prediction effect, though diminished, remains influential, as the study observed an interaction effect between RC semantics and classifier-noun agreement. When both the classifier and the RC semantics predict the same type of noun, the prediction effect is stronger, leading to more efficient processing. However, when predictions from different cues conflict, processing becomes more challenging. This finding is consistent with Kwon et al. (2017), which explored classifiers' prediction effects in a highly semantically constraining context provided by a relative clause. Their study suggests that readers pre-activate semantic features of the upcoming noun when they encounter the relative clause and the classifier, comparing these pre-activated features from both sources. When there is incongruence between them, an N400 was elicited, indicating that even in highly constraining contexts, the classifier's prediction effect is also activated.

On the other hand, classifier-noun agreement also influences comprehension question decisions in the offline processing phase. Experiment 1 demonstrated a significant increase in the selection of NP_high as the RC attachment site when there was agreement between CL (animals) and NP_high (animals). Conversely, there was a significant decrease in the selection of NP_high as the attachment site when there was agreement between CL (people) and NP_low (people), indicating a robust classifier-noun agreement effect in manipulating RC attachment preferences in offline phases.

Therefore, classifier-noun agreement can be used to resolve RC attachment ambiguities in both online and offline processing phases, influencing readers to generate either high-attachment or low-attachment preferences. With this in mind, we propose that classifier-noun agreement can function similarly to morphosyntactic agreement (e.g., gender or number) in Indo-European languages in disambiguating RC attachment, allowing us to observe attachment preferences during both the online and offline phases (Fernández, 2002). This will be discussed further in the next section.

4.2 Relative clause attachment preferences

To address the research question concerning the preference for relative clause attachment in Mandarin Chinese, our study found a low attachment during the early stage of online processing. This early low-attachment was prominent in the critical region of NP_low and its immediate spill-over region, de2. This initial low attachment can be explained by the effect of incremental processing, as observed in Japanese (Kamide and Mitchell, 1997, 1999).

Furthermore, our study suggests a shift in attachment preference during the later stage of online processing due to discourse or post-syntactic information. Specifically, a high-attachment preference surfaced in the spill-over regions of NP_high, namely ADV_main and VP_main. This shift indicated a transition from the initial low-attachment to a late-stage high-attachment preference, a pattern also observed in other languages such as Spanish (Fernández, 2002), Italian (de Vincenzi and Job, 1993), and Japanese (Kamide and Mitchell, 1997). A similar pattern was evident in the study conducted by Kamide and Mitchell (1997) in Japanese, where they manipulated pragmatic compatibility between the RC and potential attachment sites. Their findings, based on a self-paced reading task segmented into seven regions (RC/NP1-Low/Gen/NP2-High/Acc/Matrix subject/Final), demonstrated significantly shorter reading times in the regions associated with NP1-Low and Gen in the forced low-attachment condition compared to the forced high-attachment condition. Notably, the Final region in their study exhibited substantially shorter reading times in the forced high-attachment condition than in the forced low-attachment condition. These results suggested an initial low attachment when encountering the first potential attachment site (NP1-Low), followed by a shift to high attachment toward the end of the sentence, similar to the findings in our present study.

However, our findings contrast with those of Shen (2006), who observed a low-attachment preference in both the late online processing phase and the offline phase. Shen (2006) employed a self-paced reading task and an offline questionnaire to investigate RC attachment preferences in Mandarin. She manipulated the semantic compatibility between the RC and the two possible attachment sites, resulting in three conditions: ambiguous, forced-high, and forced-low conditions. The self-paced reading results indicated no significant differences between the three conditions regarding reading times at the low attachment site (NP_low), suggesting the absence of attachment preferences in the early online processing phase. However, at the high-attachment site (NP_high), reading times in the forced-high condition were significantly longer than those in the forced-low and ambiguous conditions, indicating a low-attachment preference in the late online processing phase. Furthermore, this low-attachment preference persisted into the offline phase, as demonstrated by the offline questionnaire responses (71.88% for low-attachment response).

In the following, we discuss three potential discourse/post-syntactic factors that may contribute to the observed shift to high-attachment preference in our study. First, we consider the semantic effect. In our materials, NP_low consistently referred to people while NP_high invariably denoted animals. One possibility for the observed high attachment can be attributed to people preferring animals over human beings for RC attachment, prompting a tendency to attach RCs to the distant NP_high during late online processing and offline phases, thus resulting in high attachment preferences. However, this possibility is unlikely since human nouns are more prominent in the animacy hierarchy and should be a stronger attractor for modification and attention (Hawkins, 1994). Furthermore, in constructing our stimuli, we made sure the actions described in the RCs, such as “drinking water,” “sleeping,” and “taking a rest,” are equally compatible with both animals and human beings, as was evidenced by our norming test results.

Another potential factor is the frequency effect. In both experiments, human nouns (NP_low) had higher frequencies than animal nouns (NP_high). Experiment 1 showed a marginally significant difference between NP_low (M = 0.07; SD = 0.12) and NP_high (M = 0.03; SD = 0.04) (t = 1.94, p = 0.07). Experiment 2 found a significant difference between NP_low (M = 0.07; SD = 0.11) and NP_high (M = 0.02; SD = 0.03) (t = 2.31, p < 0.05). However, since we are comparing the reading times at the NP_low region or the NP_high region across different conditions, rather than directly comparing the reading times between the NP_low and NP_high regions, frequency should not be a concern for interpreting the results. In fact, our findings show a tendency of high attachment, which is to attach the RC to a word of lower frequency. This is the opposite of what a typical frequency effect may predict (i.e., presumably words of higher frequencies may be stronger attractors).

Lastly, we turn to the potential length effect, a phenomenon posited by the Implicit Prosody Hypothesis (Fodor, 2002; Jun, 2003), which suggests that the length of RCs influences attachment preferences. Long RCs in head-final nominal structures are argued to favor high attachment, forming a single chunk that modifies the complex nominal phrase as a whole (e.g., long RC de1//NP_low de2 NP_high), while short RCs tend to group with their adjacent noun phrase, generating a preference for low attachment (e.g., short RC de1 NP_low//de2 NP_high) (Fodor, 2002; Jun, 2003; Hemforth et al., 2015). In the current study, the RCs are longer than those in two other studies on Chinese RC attachment preferences (Shen, 2006; Kwon et al., 2019)⁵ as they ranged from 7 to 12 characters (Experiment 1: M = 9.63, SD = 1.24; Experiment 2: M = 9.46, SD = 1.14). Additionally, Mandarin RCs are head-final; language users, therefore, encounter the embedded relative clause before the complex head noun is reached (Lin, 2019). Therefore, our study revealed a tendency to high-attachment driven by the long RC effect around the NP_high region, which affects the offline questionnaire stage.

However, our findings appear to contradict the production results in Kwon et al. (2019), who conducted a sentence completion task and concluded that Mandarin exhibits a preference for low attachment in the offline stage. It is noteworthy that Kwon et al. (2019) acknowledged that “almost all responses were descriptive relative clauses (78.4%)” (p. 5). In essence, most RCs generated by participants were simple attributive adjectival phrases of short length, such as qinlao de nongfu de nongchang “the farm of the farmer who is diligent,” where the RC length is three syllables. Therefore, their discovery of low-attachment preference aligns with the Implicit Prosody Hypothesis, possibly due to the short length of the RCs in their study (Fodor, 2002; Jun, 2003).

In conclusion, we explored the complexity of relative clause attachments in Mandarin Chinese, examining the influence of linguistic cues on comprehenders' attachment preferences through two self-paced reading experiments. Our results demonstrate that classifier-noun agreement in Mandarin effectively resolves RC attachment ambiguities, influencing both online and offline processing phases. This study aligns with previous research using morphosyntactic dependencies to disambiguate relative clause attachment. The findings show that classifier-noun agreement, by pre-activating semantic features of the upcoming noun, guides comprehenders' expectations and enhances processing efficiency. While the classifier's prediction effect is significant, RC semantics can overshadow it in contexts with strong disambiguating cues. However, the classifier's prediction effect remains influential, especially when congruent with RC semantics, leading to more efficient processing. This suggests that classifier-noun agreement functions similarly to morphosyntactic agreement in Indo-European languages, offering insights into attachment preferences during both online and offline phases. In addition, our study reveals a dynamic pattern of RC attachment preferences in Mandarin Chinese, characterized by an early low-attachment during early online processing, transitioning to a high-attachment preference in later online and offline stages, particularly favoring NP_high as the attachment site.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: OSF platform “Frontiers_RC Attachment” https://osf.io/dtgjc/?view_only=b574c2aa1c5748d0828fee586babf680.

Ethics statement

The studies involving humans were approved by Indiana University Human Subjects and Institutional Review Boards. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their informed consent to participate in this study.

Author contributions

FL: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Validation, Visualization, Writing – original draft. C-JL: Conceptualization, Formal analysis, Methodology, Project administration, Supervision, Writing – review & editing.

Funding

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This research was supported by C-JL's research fund and the Open Access Article Publishing Fund from the IU Libraries and the Office of the Vice Provost for Research of Indiana University Bloomington.

Acknowledgments

We are grateful to all our participants and members of the Language and Cognition Laboratory for their valuable feedback and comments. ChatGPT version 4o, a language model developed by OpenAI, has been used in the editing process to improve the readability and language of the manuscript.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

1. ^The numbers of test sentences in Experiment 1 (k = 24) and Experiment 2 (k = 28) were determined by referencing previous studies on RC attachment that employed similar methodologies (e.g., Kamide and Mitchell, 1999; Fernández, 2002; Shen, 2006; Kwon et al., 2019).

2. ^OR, Odd Ratios.

3. ^CI: 95% Confidence intervals for the coefficients.

4. ^d, Cohen's d.

5. ^Both Shen (2006) and Kwon et al. (2019) did not control RC lengths in their studies. Based on the three examples provided in the article, RC lengths in the second rating task are 4, 7, and 7 characters in Kwon et al. (2019). In Shen (2006), the RC lengths ranged from 5 to 14 characters (M = 7.64; SD = 1.89), with 87.5% of RCs having fewer than 10 characters.

References

Ahrens, K., and Huang, C. (2016). “Classifiers,” in A Reference Grammar of Chinese Reference Grammars, eds. C. Huang and Shi, D. (Cambridge: Cambridge University Press), 169–198. doi: 10.1017/CBO9781139028462.008

Crossref Full Text | Google Scholar

Akaike, H. (1998). “Information theory and an extension of the maximum likelihood principle,” in Selected Papers of Hirotugu Akaike, eds. E. Parzen, K. Tanabe, and Kitagawa, G. (New York: Springer), 199–213. doi: 10.1007/978-1-4612-1694-0_15

Crossref Full Text | Google Scholar

Barr, D. J., Levy, R., Scheepers, C., and Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: keep it maximal. J. Memory Lang. 68, 255–278. doi: 10.1016/j.jml.2012.11.001

PubMed Abstract | Crossref Full Text | Google Scholar

Box, G. E. P., and Cox, D. R. (1964). An analysis of transformations. J. Royal Stat. Soc. Series B (Methodological) 26, 211–252. doi: 10.1111/j.2517-6161.1964.tb00553.x

Crossref Full Text | Google Scholar

Brysbaert, M., and Mitchell, D. C. (1996). Modifier attachment in sentence processing: Evidence from Dutch. Quart. J. Exp. Psychol. 49A, 664–695. doi: 10.1080/713755636

PubMed Abstract | Crossref Full Text | Google Scholar

Chan, S. H. (2019). An elephant needs a head but a horse does not: an ERP study of classifier-noun agreement in Mandarin. J. Neuroling. 52:100852. doi: 10.1016/j.jneuroling.2019.100852

Crossref Full Text | Google Scholar

Chou, C. J., Chang, C. T., Tsai, J. L., and Lee, C. Y. (2015). “Interplay between semantic and syntactic information in Chinese Classifier-noun agreement: An ERP comparison,” in Poster at the Seventh Annual Neurobiology of Language Conference (SNL 2015) (Chicago, IL).

Google Scholar

Chow, W.-Y., and Chen, D. (2020). Predicting (in)correctly: listeners rapidly use unexpected information to revise their predictions. Lang. Cognit. Neurosci. 35, 1149–1161. doi: 10.1080/23273798.2020.1733627

Crossref Full Text | Google Scholar

Cotter, B. T., and Ferreira, F. (2014). The relationship between working memory capacity, bilingualism, and ambiguous relative clause attachment. Mem. Cognit. 1–18. doi: 10.3758/s13421-024-01561-4

PubMed Abstract | Crossref Full Text | Google Scholar

Cuetos, F., and Mitchell, D. C. (1988). Cross-linguistic differences in parsing: restrictions on the use of the Late Closure strategy in Spanish. Cognition 30, 73–105. doi: 10.1016/0010-0277(88)90004-2

PubMed Abstract | Crossref Full Text | Google Scholar

Daneman, M., and Carpenter, P. A. (1980). Individual differences in working memory and reading. J. Verbal Learning Verbal Behav.19, 450–466. doi: 10.1016/S0022-5371(80)90312-6

Crossref Full Text | Google Scholar

de Leeuw, J. R., Gilbert, R. A., and Luchterhandt, B. (2023). jsPsych: Enabling an open-source collaborative ecosystem of behavioral experiments. J. Open Source Softw. 8:5351. doi: 10.21105/joss.05351

Crossref Full Text | Google Scholar

de Vincenzi, M., and Job, R. (1993). Some observations on the universality of the late-closure strategy. J. Psycholinguist. Res. 22, 189–206. doi: 10.1007/BF01067830

Crossref Full Text | Google Scholar

Erbaugh, M. (2006). “Chinese classifiers: Their use and acquisition.” in The handbook of East Asian Psycholinguistics, eds. P. Li, H. T. Li, E. Bates, and J. L. Tzeng (Cambridge: Cambridge University Press), 39–51. doi: 10.1017/CBO9780511550751.005

Crossref Full Text | Google Scholar

Federmeier, K. D., and Kutas, M. (1999). A rose by any other name: long-term memory structure and sentence processing. J. Mem. Lang. 41, 469–495. doi: 10.1006/jmla.1999.2660

Crossref Full Text | Google Scholar

Fernández, E. M. (2002). “Relative clause attachment in bilinguals and monolinguals,” in Bilingual Sentence Processing, eds. R. R. Heredia and J. Altarriba (Amsterdam, NL: North Holland Elsevier Science Publishers), 187–216.

PubMed Abstract | Google Scholar

Fodor, J. D. (2002). “Prosodic disambiguation in silent reading,” in Proceedings of the North East Linguistic Society (NELS), ed. M. Hirotani (Amherst, MA: GLSA Publications), 112–132.

Google Scholar

Grüter, T., Lau, E., and Ling, W. (2019). How classifiers facilitate predictive processing in L1 and L2 Chinese: the role of semantic and grammatical cues. Lang. Cognit. Neurosci. 35, 221–234. doi: 10.1080/23273798.2019.1648840

Crossref Full Text | Google Scholar

Harding, E. E., Sammler, D., and Kotz, S. A. (2019). Attachment preference in auditory German sentences: Individual differences and pragmatic strategy. J. Front. Psychol. 10, 1–10. doi: 10.3389/fpsyg.2019.01357

PubMed Abstract | Crossref Full Text | Google Scholar

Hawkins, J. A. (1994). A Performance Theory of Order and Constituency. Cambridge: Cambridge University Press.

Google Scholar

Hemforth, B., Fernández, S., Clifton, C., Frazier, L., Konieczny, L., and Walter, M. (2015). Relative clause attachment in German, English, Spanish and French: effects of position and length. Lingua 166, 43–64. doi: 10.1016/j.lingua.2015.08.010

Crossref Full Text | Google Scholar

Hemforth, B., Konieczny, L., and Scheepers, C. (1994). “Principle based and probabilistic approaches to human parsing: how universal is the human language processor?” in Paper presented at the Verarbeitung natürlicher Sprache Conference (KONVENS'94).

Google Scholar

Hopp, H., and Lemmerth, N. (2018). Lexical and syntactic congruency in L2 predictive gender processing. Stud. Second Lang. Acquisit. 40, 171–199. doi: 10.1017/S0272263116000437

PubMed Abstract | Crossref Full Text | Google Scholar

Hsu, C. N., Tsai, S., Yang, C. L., and Chen, J. (2014). Processing classifier–noun agreement in a long distance: an ERP study on Mandarin Chinese. Brain Lang. 137, 14–28. doi: 10.1016/j.bandl.2014.07.002

PubMed Abstract | Crossref Full Text | Google Scholar

Huang, Z., Feng, C., and Qu, Q. (2023). Predicting coarse-grained semantic features in language comprehension: evidence from ERP representational similarity analysis and Chinese classifier. Cereb. Cortex 33, 8312–8320. doi: 10.1093/cercor/bhad116

PubMed Abstract | Crossref Full Text | Google Scholar

Jun, S. A. (2003). Prosodic phrasing and attachment preferences. J. Psycholinguist. Res 32, 219–249. doi: 10.1023/A:1022452408944

PubMed Abstract | Crossref Full Text | Google Scholar

Kamide, Y., Altmann, G. T. M., and Haywood, S. L. (2003). The time-course of prediction in incremental sentence processing: evidence from anticipatory eye movements. J. Mem. Lang. 49, 133–156. doi: 10.1016/S0749-596X(03)00023-8

PubMed Abstract | Crossref Full Text | Google Scholar

Kamide, Y., and Mitchell, D. C. (1997). Relative clause attachment: non-determinism in Japanese parsing. J. Psycholinguist. Res. 26, 247–254. doi: 10.1023/A:1025017817290

Crossref Full Text | Google Scholar

Kamide, Y., and Mitchell, D. C. (1999). Incremental pre-head attachment in japanese parsing. Lang. Cognit. Proc. 14, 631–662. doi: 10.1080/016909699386211

Crossref Full Text | Google Scholar

Kuperberg, G. R., and Jaeger, T. F. (2016). What do we mean by prediction in language comprehension? Lang. Cognit. Neurosci 31, 32–59. doi: 10.1080/23273798.2015.1102299

PubMed Abstract | Crossref Full Text | Google Scholar

Kwon, N., Ong, D., Chen, H., and Zhang, A. (2019). The role of animacy and structural information in Relative Clause Attachment: Evidence from Chinese. Front. Psychol. 10:1576. doi: 10.3389/fpsyg.2019.01576

PubMed Abstract | Crossref Full Text | Google Scholar

Kwon, N., Sturt, P., and Liu, P. (2017). Predicting semantic features in Chinese: evidence from ERPs. Cognition 166, 433–446. doi: 10.1016/j.cognition.2017.06.010

PubMed Abstract | Crossref Full Text | Google Scholar

Lee, D., and Kweon, S. (2004). A sentence processing study of relative clause in Korean with two attachment sites. Discourse Cogn. 11, 125–140.

Google Scholar

Lee, S. Y. (2021). The Effect of Honorific Affix on Processing of an Attachment Ambiguity. Japanese/Korean Linguist. 28, 1–10.

Google Scholar

Lee, S. Y., and Yoo, M. Y. (2023). Honorific agreement in Korean: retrieval processing and predictive processing. Linguist. Res. 40, 387–407. doi: 10.17250/KHISLI.40.3.202312.002

PubMed Abstract | Crossref Full Text | Google Scholar

Lin, C. J. C. (2019). “Chinese psycholinguistics: A typological overview,” in The Routledge Handbook of Chinese Applied Linguistics, eds. R. C. R. Huang, J. S. Zhuo, and B. Meisterernst (London: Routledge), 773–785. doi: 10.4324/9781315625157-48

Crossref Full Text | Google Scholar

Lin, C. J. C., and Bever, T. G. (2011). “Garden path and the comprehension of head-final relative clauses,” in Processing and Producing Head-Final Structures, eds. H. Yamashita, Y. Hirose, and J. L. Packard (New York, NY: Springer), 277–297. doi: 10.1007/978-90-481-9213-7_13

Crossref Full Text | Google Scholar

Linck, J., and Cunnings, I. (2015). The utility and application of mixed effects models in second language research. Lang. Learn. 65, 185–207. doi: 10.1111/lang.12117

Crossref Full Text | Google Scholar

Luce, R. D. (1986). Response Times: Their Role in Inferring Elementary Mental Organization. New York: Oxford University Press.

PubMed Abstract | Google Scholar

Mahmoodi, M. H., Sheykholmoluki, H., Zoghipaydar, M. R., and Shahsavari, S. (2022). Working memory capacity and relative clause attachment preference of Persian EFL learners: does segmentation play any role? J. Psycholinguist. Res. 51, 683–706. doi: 10.1007/s10936-021-09825-9

PubMed Abstract | Crossref Full Text | Google Scholar

Marefat, H., Samadi, E., and Yaseri, M. (2015). Semantic priming effect on relative clause attachment ambiguity resolution in L2. Appl. Res. Engl. Lang. 4, 78–95. doi: 10.22108/are.2015.15504

Crossref Full Text | Google Scholar

Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., and Bates, D. (2017). Balancing type I error and power in linear mixed models. J. Mem. Lang. 94, 305–315. doi: 10.1016/j.jml.2017.01.001

Crossref Full Text | Google Scholar

Mendelsohn, A., and Pearlmutter, N. (1999). “Individual differences in relative clause attachment preferences,” in Poster Presented at the CUNY Conference on Human Sentence Processing (New York, NY).

Google Scholar

Mitchell, D. C., and Cuetos, F. (1991). “The origins of parsing strategies,” in Current Issues in Natural Language Processing, ed. C. Smith (Austin: Center for Cognitive Science, University of Texas) 1–12.

Google Scholar

Omaki, A. (2005). Working Memory and Relative Clause Attachment in First and Second Language Processing [Doctoral dissertation]. Honolulu: University of Hawaii at Manoa.

Google Scholar

Özge, D., Kornfilt, J., Maquate, K., Küntay, A. C., and Snedeker, J. (2022). German-speaking children use sentence-initial case marking for predictive language processing at age four. Cognition 221:104988. doi: 10.1016/j.cognition.2021.104988

PubMed Abstract | Crossref Full Text | Google Scholar

Özge, D., Küntay, A., and Snedeker, J. (2019). Why wait for the verb? Turkish speaking children use case markers for incremental language comprehension. Cognition 183, 152–180. doi: 10.1016/j.cognition.2018.10.026

PubMed Abstract | Crossref Full Text | Google Scholar

Pickering, M. J., and Gambi, C. (2018). Predicting while comprehending language: a theory and review. Psychol. Bull. 144, 1002–1044. doi: 10.1037/bul0000158

PubMed Abstract | Crossref Full Text | Google Scholar

Pickering, M. J., and Garrod, S. (2013). An integrated theory of language production and comprehension. Behav. Brain Sci. 36, 329–347. doi: 10.1017/S0140525X12001495

PubMed Abstract | Crossref Full Text | Google Scholar

R Core Team (2023). R: A Language and Environment for Statistical Computing. Vienna: R Foundation for Statistical Computing. Available at: https://www.R-project.org/ (accessed December 20, 2023).

Google Scholar

Roberts, L., and Felser, C. (2011). Plausibility and recovery from garden paths in second language sentence processing. Appl. Psycholinguist. 32, 299–331. doi: 10.1017/S0142716410000421

Crossref Full Text | Google Scholar

Shen, X. (2006). Late Assignment of Syntax Theory: Evidence from Chinese and English. [Doctoral dissertation]. Exeter: University of Exeter.

Google Scholar

Swets, B., Desmet, T., Hambrick, D. Z., and Ferreira, F. (2007). The role of working memory in syntactic ambiguity resolution: a psychometric approach. J. Exp. Psychol. 136, 64–81. doi: 10.1037/0096-3445.136.1.64

PubMed Abstract | Crossref Full Text | Google Scholar

Traxler, J. M. (2007). Working memory contributions to relative clause attachment processing: a hierarchical linear modeling analysis. Mem. Cognit. 35, 1107–1121. doi: 10.3758/BF03193482

PubMed Abstract | Crossref Full Text | Google Scholar

Traxler, J. M. (2009). A hierarchical linear modeling analysis of working memory and implicit prosody in the resolution of adjunct attachment ambiguity. J. Psycholinguist. Res. 38, 491–509. doi: 10.1007/s10936-009-9102-x

PubMed Abstract | Crossref Full Text | Google Scholar

Wang, L., Kuperberg, G., and Jensen, O. (2018). Specific lexico-semantic predictions are associated with unique spatial and temporal patterns of neural activity. Elife 7:e39061. doi: 10.7554/eLife.39061.019

PubMed Abstract | Crossref Full Text | Google Scholar

Wu, F., Kaiser, E., and Vasishth, S. (2018). Effects of early cues on the processing of Chinese relative clauses: evidence for experience-based theories. Cogn. Sci. 42, 1101–1133. doi: 10.1111/cogs.12551

PubMed Abstract | Crossref Full Text | Google Scholar

Zagar, D., Pynte, J., and Rativeau, S. (1997). Evidence for early-closure attachment on first-pass reading times in French. Q. J. Exp. Psychol. 50, 421–438. doi: 10.1080/713755715

Crossref Full Text | Google Scholar

Zhou, X., Jiang, X., Ye, Z., Zhang, Y., Lou, K., and Zhan, W. (2010). Semantic integration processes at different levels of syntactic hierarchy during sentence comprehension: an ERP study. Neuropsychologia 48, 1551–1562. doi: 10.1016/j.neuropsychologia.2010.02.001

PubMed Abstract | Crossref Full Text | Google Scholar